You Bought a Tool. The Value Was in the Workflow You Didn't Redesign.

Roughly 84% of organizations that bought AI never redesigned a single workflow around it, according to Gartner figures relayed by Jeff Gothelf (via Computerworld, June 2026). They licensed the model, wired it into the existing process, and waited for the return. The return never arrived. About 80% of those companies also cut headcount with no measurable link to ROI, which is the predictable result of treating a tool purchase as a transformation.

That number is the whole story. The failure of enterprise AI is rarely a failure of the model. It is a failure to change the work the model was supposed to improve.

The Tool Was Never the Deliverable

Gothelf’s argument starts from a familiar pattern. A team buys a capable model, integrates it at the edge of a process that was designed for humans doing the work by hand, and is surprised when output barely moves. The integration is real. The redesign is missing. The AI sits inside a workflow whose handoffs, approvals, and quality checks all assume the old division of labor.

This is why so many pilots stall. A large share of enterprise AI pilots never pass proof-of-concept, per an MIT study cited secondhand by Gothelf (via Forbes). The PoC works because someone hand-holds it through a clean demo path. Production breaks because production is the messy workflow nobody redesigned. The last mile is cultural and organizational, not technical. You can buy a frontier model in an afternoon. You cannot buy the redesigned process, the new escalation rules, or the team’s trust in the output.

We have written before about the distance between buying AI and operating it (see The ROI gap in AI-native organizations). The 84% figure puts a number on the cause. The tool was a component. The workflow was the product, and most teams never built it.

What “Redesigning the Workflow” Actually Means

Redesigning a workflow is not adding an AI step to a flowchart. It means deciding, concretely:

Which decisions the agent makes alone, and which it routes to a human
What gets logged, so the output can be audited after the fact
Where the new quality checkpoints sit, now that the old human handoffs are gone
How exceptions escalate when the model is uncertain or wrong

Each of those is a governance decision wearing an operations costume. When you specify which decisions an agent can make without review, you are writing policy. When you decide what gets logged, you are building the audit trail. When you place the new checkpoints, you are designing oversight. The redesigned workflow and the governance layer are the same artifact, described from two angles.

This is the part that does not transfer in a slide deck. Coordination between humans and agents has to be designed deliberately (see Coordination-first AI automation). A model dropped into an unchanged process inherits all the coordination assumptions of the old one, including the ones that no longer hold once a machine is doing part of the work.

The Same Insight, Read Forward, Is the Moat

Jamin Ball of Altimeter Capital, writing in Clouded Judgement (June 2026), arrives at the workflow from the opposite direction. He is looking at where durable value accrues in the agentic era, and his answer is the orchestration layer: the system that manages, routes, and governs the agents. That layer, he argues, becomes the new system of record.

The market is already pricing this. Ball notes that the top-five SaaS companies by forward revenue multiple trade at a median of 26.4x next-twelve-months revenue, against an overall median of 2.9x. The premium sits with the companies positioned to own orchestration, not with the ones selling a single capability that any competitor can also license.

Put Gothelf and Ball side by side and they describe one thing from two ends. Gothelf shows the failure: a workflow nobody redesigned, so the AI delivers nothing. Ball shows the upside: the layer that orchestrates and governs agents becomes the asset competitors cannot easily copy. The bridge between them is direct. Redesigning the workflow is how you build the orchestration layer. The governed, redesigned process is the moat Ball is pricing and the deliverable Gothelf says most teams skipped.

A licensed model is available to everyone who can pay for it. A workflow that routes work between your people and your agents, logs every consequential decision, and enforces your escalation rules is specific to your operation. The first is a commodity. The second compounds.

Why This Matters for Measurement

If the workflow is the unit of value, then measuring the tool tells you almost nothing. A benchmark score for the model is not a measure of your throughput, your error rate, or your team’s trust in the output. Those live in the redesigned process, on a human-plus-AI team, and they are what actually moved or failed to move.

This is the same reason headcount is the wrong scoreboard. The 80% who cut staff with no ROI correlation were measuring the input they could see (people removed) instead of the output they were supposed to improve (work delivered per person on a governed team). The companies with the highest returns invested in productivity growth inside redesigned workflows, not in subtraction.

The orchestration layer Ball describes is also the measurement layer. The system that routes and governs the agents is the only place where you can see, per workflow, what the human-plus-AI team actually produced. Build the workflow and you get governance, a moat, and a scoreboard from the same investment. Skip it and you have an expensive model sitting inside a process that was never going to convert it into value.

Do This Now

Pick one workflow where you have already deployed AI and ask a single question: what did we redesign? If the answer is “we added the model to the existing steps,” you are in the 84%. Before buying another seat or another tool, redesign that one process end to end. Decide which decisions the agent owns, what gets logged, where the new checkpoints sit, and how exceptions escalate. That redesigned, governed workflow is the deliverable you thought you were buying. It is also the start of the moat, and the only surface on which you can honestly measure what your AI is doing.

This analysis synthesizes How Product Management Can Fix Your AI Integration Problems (Jeff Gothelf, June 2026), citing Gartner (via Computerworld) and an MIT study (via Forbes), and Clouded Judgement 6.19.26: Workflows are King (Altimeter Capital, Jamin Ball, June 2026).

Victorino Group helps teams redesign the workflow into a governed, measurable system of record, not just install another tool. Let’s talk.